deepboof.backward

Interface NumericalGradient<T extends deepboof.Tensor<T>>

All Known Implementing Classes:

NumericalGradient_F64

public interface NumericalGradient<T extends deepboof.Tensor<T>>

Given a Function implementations of this interface will compute the gradient of its inputs and parameters. Numerical differentiation is done using a symmetric sample, e.g. dx = [f(x+T)-f(x-T)]/T

Method Summary

Modifier and Type	Method and Description
void	configure(double T) Overrides default settings for computing numerical gradient.
void	differentiate(T input, java.util.List<T> parameters, T dout, T gradientInput, java.util.List<T> gradientParameters) Performs numerical differentiation to compute the gradients of input and parameters.
void	setFunction(deepboof.Function<T> function) Sets the function which will be differentiated and other parameters.

Method Detail

configure

void configure(double T)

Overrides default settings for computing numerical gradient.

Parameters:

T - Sampling distance used for numerical differentiation

setFunction

void setFunction(deepboof.Function<T> function)

Sets the function which will be differentiated and other parameters. Function.initialize(int...) should have already been called.

Parameters:

function - The function which is to be differentiated

differentiate

void differentiate(T input,
                   java.util.List<T> parameters,
                   T dout,
                   T gradientInput,
                   java.util.List<T> gradientParameters)

Performs numerical differentiation to compute the gradients of input and parameters. When numerical differentiation is being performed input and parameters will be modified and then returned to their original state.

Parameters:

input - The same input tensor which was passed in during the forward pass.

parameters - The same parameters which was passed in during the forward pass.

dout - Derivative of output, computed from next layer.

gradientInput - Storage for gradient of input

gradientParameters - Storage for gradients of parameters